<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-2126">
  <Title>Anaphorie Reference to Events and Actions: A Representation and its Advantages</Title>
  <Section position="3" start_page="602" end_page="602" type="metho">
    <SectionTitle>
2 Discourse Models
</SectionTitle>
    <Paragraph position="0"> An overall concern of this work is to describe how communication is done via discourse. A piece of discourse is a set of sentences or pieces of text that are spoken by one or more speakers. When we talk about discourse, we usually indicate the fact that those sentences in the discourse are connected in some way that makes them comprehensible and coherent. Speakers do so by attempting to get their listeners to construct an appropriate model: a discourse model. A discourse model is viewed as containing representations of entities, along with their properties and relations they participate in \[Webber 1982, Helm 1982, Kamp 1984\]. The key, then, in successful communication is for the speaker to transmit as much information about those representations of the entities, their properties and relations to the listener. Usually, this cannot be done in a single utterance, so it requires reference back to things mentioned previously. The speaker refers to a particular representation of a person, object, action or event in his or her discourse model and he or she can do so by using a pronoun, a definite noun phrase, or a form such as do it, do that. The speaker expects the listener to (i) have in his or her discourse model a representation for those objects, actions or events (or to be able to create one with minimal effort), and (ii) make the link between the referent 1 and the representation of the entity.</Paragraph>
    <Paragraph position="1"> Discourse entities can represent objects, sets, actions, events, and facts which are viewed as individuals. The referent of a definite pronoun is then an entity in the speaker's discourse model which, given the goal of the communication, is expected to exist XReseareh in linguistics has defined the problem of anaphom as that of finding the antecedent of an anaphoric expression, that is, finding the word or phrase to which the anaphor is linked. In this work, the problem of anaphom is defined not solely as that of identifying the antedecent but how the referent is found. We consider the antecedent to be the linguistic text which gives rise to an entity, as compared to a referent which is a non-linguistic entity.</Paragraph>
    <Paragraph position="2"> also in the listener's model, hi the next section, we describe how discourse entities describing events or actions can be represented in a discourse model in order to indicate what they make available for future reference.</Paragraph>
  </Section>
  <Section position="4" start_page="602" end_page="603" type="metho">
    <SectionTitle>
3 Formalizing the Representation
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="602" end_page="602" type="sub_section">
      <SectionTitle>
3.1 Characterization
</SectionTitle>
      <Paragraph position="0"> First, let us consider what characterizes actions or events. In general, each sentence gives rise to an event. Similarly, an action is considered to be part of that event. As a way to establish the differences linguistically, we describe actions as being characterized by the predicate of the sentence (the verb phrase) and events as being characterized hy the whole sentence. Therefore, we consider actions to be part of events. This notion is similar to Jackendoff's who notes that &amp;quot;sentences that express \[ACTIONS\] are a subset of those that express \[EVENTS\]&amp;quot; \[Jaekendoff 1983\] (p. 179). He proposes introducing &amp;quot;what happened was&amp;quot; as a diagnostic for \[EVENTS\] and &amp;quot;what x did&amp;quot; as a diagnostic for \[ACTIONS\]. So we have  (3) What happened was that the pig ran away. ~ EVENT and (4) What Fled did was run away. ~ ACTION  Withiu this view, &amp;quot;an \[EVENT\] that is also an \[ACTION\] involves a character with a special role- the one who is performing the \[ACTION\],&amp;quot; called the \[ACTOR\] (p. 180).</Paragraph>
      <Paragraph position="1"> Consider for example: (5) Mary deleted the region.</Paragraph>
      <Paragraph position="2"> Here, the action of deleting, which occurred in some past time, was performed by Mary and it was done on the region. 2 The 'delete' action relates Mary and the region.</Paragraph>
      <Paragraph position="3"> Events correspond to the conjunction of the action predicate with other predicates described in the sentence such as time, place, along with the agent performing the action. Tense and aspect usually indicate the sense of time in an event. The tense indicates present, past, or future and the aspect indicates the perfect (completion) and progressive (continuous) forms. In general, they are indicated syntactically by certain verbs and/or the auxiliaries. For instance, the auxiliary verb gives us a sense of whether the event is occurring (presen0, occurred (past), will occur (future) or is just hypothesized-as well as whether an event is continuous or it has been completed. 3 Events and actions usually have associated with them a time interval. This time interval can have a well defined beginning and/or end. There may also exist events which occur as points, that is, their beginning and end times are the same as well as those in which there is no end or no beginning. These are among the issues that provide us with information about the event or the action. There is a vast literature of these subjects including works by \[Vendler 1967, Mourelatos 1978, Dowty 1986, Hinrichs 1986, Moens et al. 1987, Passonneau 1987\] to mention some.</Paragraph>
      <Paragraph position="4"> 2In EMACS, a region is a block of text defined by the mark at one end and the cursor at the other end. More precisely, it is the area of the buffer between the dot and the mark.</Paragraph>
      <Paragraph position="5"> aEnglish has two divisions for aspect: (i) Progressive (verb to be and the verbal form -ing) and non-Progressive, and (it) Perfect (verb to have and Past Participle) and mm-PerfecL English also has a separate Habitual (which occurs only in the past tense), using the auxiliary form used to. The habitual can be replaced by its non-habitual equivalent, i.e. the non-habitual does not exclude habitual meaning \[Comrie 1976\].</Paragraph>
      <Paragraph position="6"> To summarize, our characterization of events is based on individual sentences. Each event includes an action predicate which, along with other predicates describing tense, aspect, and the agent performing the action constitutes the particular event.</Paragraph>
    </Section>
    <Section position="2" start_page="602" end_page="603" type="sub_section">
      <SectionTitle>
3.2 Representation of Events
</SectionTitle>
      <Paragraph position="0"> Given our characterization of events and actions as described before, and assuming a syntactic representation for each sentence, the next step is to describe the kind of (semantic) representation we use to characterize those events. Many representations for events have been proposed to handle similar issues as we describe here \[Davidson 1967, Moore 1981, Neale 1987\]. The representation we have chosen is AProlog, a fortnaI logical representation described by \[Miller et al. 1986\]. This approach allows us to represent entities describing events and actions in a logical and complex way. In this representation, the logical form describes the events that characterize the sentence (as compared to a proposition or individual), therefore allowing subsequent references to any one of the events described by the sentence. For example, (6) Mary types control-W.</Paragraph>
      <Paragraph position="1"> which is usually represented as type(Mary, control-W), will get mapped onto the following semantic representation: 3v\[type( Mary, control-W, v)\], which posits an individual typing event following \[Davidson 1967\]. From this semantic representation, we obtain the action described by the sentence which corresponds to its predicate (e.g. 'type control-W' in this example). This action can be represented in terms of )~ predicates as in, Ax.\[tvpe( x, control-W, e)\].</Paragraph>
      <Paragraph position="2"> The sentence characterizes an event description which is represented as</Paragraph>
      <Paragraph position="4"> meaning that E is an entity describable as 'the event in which Mary types control-W'. This is the entity associated with the event description characterized by the sentence. Present(e) means that the event occurs now. 4 As a way of illustrating the representation, consider an example:  (7) John was shot in broad daylight in Philadelphia.</Paragraph>
      <Paragraph position="5"> which can be represented as follows: during(daylight_hours, e) A in(Philadelphia, e) A past(e)\].</Paragraph>
      <Paragraph position="6">  4\[Hobbs 1985\] describes a similar approach by introducing what he calls a &amp;quot;nominalization&amp;quot; operator ' : p ~ p': Corresponding to every n-ary predicate p there will be an (n+ 1)ary predicate p' whose first argument can be thought of as the condition that holds when p is true of the subsequent arguments. Thus if run(1) means that John runs, run'(E,l) means that E is a running event by John, or John's running ...... The effect of this notational maneuver is to provide handles by which various predications can be grasped by higher predications (p. 62).  Here, E(r)-~ can be described as 'the event in which John was shot in broad.daylight in Philadelphia') E(r)4 corresponds to the real world event description in which John was shot which took place at a time during the daylight_hours, in Philadelphia. z is the person who did the shooting, who is not explicitly mentioned in the sentence. Here, past(e) is defined as past(e ) --C/, time(e) &lt; now.</Paragraph>
      <Paragraph position="7"> So E(r)-~ makes a predication about the event in which John was shot and the details that form part of that predication. Given this representation of the event, the next issue is how to generate references to that event. That is, we want to be able to generate references to (7) as in:  (8) John was shot in broad daylight in Philadelphia. It happened at 10 am.</Paragraph>
      <Paragraph position="8"> it = the event in which John was shot in broad daylight in Philadelphia.</Paragraph>
      <Paragraph position="9"> (9) John was shot in broad daylight in Philadelphia. It shocked me very much.</Paragraph>
      <Paragraph position="10"> it = the event in which John was shot in broad daylight in Philadelphia.</Paragraph>
      <Paragraph position="11"> (10) John was shot in broad daylight in Philadelphia. It has never happened before.</Paragraph>
      <Paragraph position="12"> it = an event in which someone was shot in broad daylight in Philadelphia.</Paragraph>
      <Paragraph position="13"> (11) John was shot in broad daylight in Philadelphia. AI Pacino had done it last year in a warm summer evening in New York City.</Paragraph>
      <Paragraph position="14"> it = shoot someone/John.</Paragraph>
      <Paragraph position="15"> (12) John was shot in broad daylight in Philadelphia. That would never happen in BogotL that = an event in which someone was shot in broad daylight.  As seen in these examples, the referents of the pronouns may have been introduced explicitly. That is, the anaphoric expression in the second sentence can refer to the specific event as characterized by the sentence or to a less specific event than the one described by the sentence. Therefore, what we need is to have access to the specific event description along with more general event descriptions of the one characterized by the sentence. This representation allows us to do so by being able to perform a generalization on the specific event description as we show in the next section.</Paragraph>
    </Section>
    <Section position="3" start_page="603" end_page="603" type="sub_section">
      <SectionTitle>
3.3 Generalization of Event Descriptions
</SectionTitle>
      <Paragraph position="0"> In general, each event represented as described in the previous section denotes a set of events. Given that there are entities corresponding to events represented in the discourse model, each of which can have one or more descriptions, we can generalize to obtain the various descriptions. Generalization is defined as follows: Given an event description E1 and an event description E2, 5As pointed out by \[Sidner 1982\], referring expressions specify discourse entities; referring expressions may co-specO~y the same discourse entity; discourse entities represent objects or events in the world and people refer to objects and events in the world when they use referring expressions.</Paragraph>
      <Paragraph position="1"> we say that E1 generalizes E2 if every event description denoted by E2 is a member of the set of event descriptions denoted by El.6 As we generalize tile representation of the event that we have obtained for each sentence, we incrementally obtain event descriptions of that event. These event descriptions are available in the discourse model for future reference. The generalization of the events fits in an ordered structure in which the most specific representation (corresponding to the description of the event characterized by the original sentence) is at the bottom and the most general at the top. So if we consider (7) again  And we can continue to generalize this last representation and obtain the following:</Paragraph>
      <Paragraph position="3"> Eft)-,: Ae.\[3x3y\[shoot( x, y, e)\]\].</Paragraph>
      <Paragraph position="4"> As specified by the generalization, every member of E(r)-~ is a member of E(r)-r. This set of generalizations can be ordered in a graph-like structure as shown in Fig. 1.</Paragraph>
      <Paragraph position="5"> In this structure, which provides a partial ordering of the events, some events are more specific than others. The structure is partially ordered, with the &lt; relation. 7 We have not generalized to 'any time t' in the !during' predicate because all events occur during some time. Note that at the very bottom of our graph we could generalize on the predicate 'shoot' so as to have an event describing &amp;quot;somebody doing something&amp;quot; which is common to all events. We have chosen not to do so in order to have access to the initial referent of the pronoun, characterized by the predicate of the event. The generalization structure provides the potential referents for the pronoun s as they appear in examples (8), (9), (10), (12), and (11). Consider for instance, the pair of sentences in (10), where the referent of the pronoun it in the second sentence is &amp;quot;an event in 6This definition is based on Millet and Nadathur's definition of subsumption (= generalization). They define it in terms of concepts as follows: &amp;quot;a concept Ca subsumes another concept (?2 if every element of the set denoted by C~ is a member of the set denoted by C1&amp;quot; lMitlet et al 1986\] (13. 6).</Paragraph>
      <Paragraph position="7"> which someone was shot in broad daylight in Philadelphia.,&amp;quot;. This event description is E(7)-4 which is available in the generalization graph. The same is true for the pair of sentences in (11) where the referent of the pronoun it is E(7).7, and the pair (12) where the referent of the pronoun that in the second sentence is E(r)-s, both of which are available thru the generalization graph.</Paragraph>
      <Paragraph position="8"> Bauerle has proposed an alternative method for dealing with event reference within the framework of Disconrse Representation Theory (DRT) \[Kamp 1984\]. As he points out, the problem is not that simple because by introducing an event-argument, the possible referents for event-anaphora are only linked to the event-token (the specific event description) and not to the type (the generalized descriptions) \[Bauerle 1988\] (p. 21). The representation outlined in this paper does provide us with access to the event-token and its generalization allows us to obtain the types.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="603" end_page="605" type="metho">
    <SectionTitle>
4 Relationship between Pronouns and their
Referents
</SectionTitle>
    <Paragraph position="0"> In addition to relating a particular pronoun and its referent, either as a specific event-token or as an event-type obtained from the generalization of the event-token, there are cases in which the pronoun is also linked to not just an'individual event but to either a sequence of events or even a particular event within a sequence (usually the last event in the sequence).</Paragraph>
    <Paragraph position="1"> As described in \[Schuster 1988\], we have studied the relationship between the pronouns and their referents, within a given text describing a sequence of events. We focused on the use and interpretation of tim pronouns it and that when referring those events and/or actions, given the representation described in tbe previous section. Initially, we were interested in identifying tile relationship between a specific pronoun such as that or it and an event as its referent. We found out that a particular pronoun could act as a referent of one of the following:  1. One event.</Paragraph>
    <Paragraph position="2"> 2. A sequence of events, the pronoun referring to the entire sequence as one.</Paragraph>
    <Paragraph position="3"> 3. The last event in a sequence.</Paragraph>
    <Paragraph position="4"> &amp;quot;4. A pair of events related by two possible relations: genera.  tion and enablement, as proposed by \[Goldman 1970\] and developed by \[Pollack 1986a\].</Paragraph>
    <Paragraph position="5"> We will describe each one of'these cases, along with some examples and relate them to the representation we introduced in section 3.  1. One event Consider the following text: (13) E: Write a simple macro that does three forward charaeters. Try that.</Paragraph>
    <Paragraph position="6">  We can represent the first sentence in the pair as described before:</Paragraph>
    <Paragraph position="8"> Here, E03}-x can be described as 'the event in which U writes a simple macro that does three forward-characters, and takes place at the present time'. U in this case is the user interacting with the expert. Then, the referent of the pronoun that in the second sentence is the event represented as E03)-l.</Paragraph>
    <Paragraph position="9">  2. A sequence of events Consider: (14) E:.To kill a region, the whole tiring at one time you  should set the mark- (esc)-M-at the top of the region, then move the cursor down to the bottom of the region and type (esc)control-K. This will kill the region.</Paragraph>
    <Paragraph position="10"> The pronoun this refers to the entire sequence of events. As in the previous case, if we represent each sentence as a Aexpression, we have access to the various descriptions of the event corresponding to each sentence and we can refer to them via the pronoun.</Paragraph>
    <Paragraph position="11">  3. Last event Consider the following example, where the pronoun it refers to the last event in the sequence: (15) E: To kill a word, you have to move tbe cursor to the beginning of that word and type control-D. Go ahead, do it.</Paragraph>
    <Paragraph position="12"> 4. Generation and Enablement In analyzing the data we'found the need for two important relationships: generation and enablement \[Pollack 1986b\]. Generation is defined as follows: If an agent performs one action and thereby, without any effort on his/her part, does another, then we can say that his/her performance of the former action &amp;quot;generated&amp;quot; the performance of the latter. For example, (16) E: Do this: set a &amp;quot;mark&amp;quot; at some point (any 01d point) by typing (esc)-M. It will say &amp;quot;mark set&amp;quot;, try it. (17) E: (esc)-M will give set-mark. Do it.</Paragraph>
    <Paragraph position="13">  In both cases, the referent(s) of the pronoun it can be either &amp;quot;setting the mark&amp;quot; or &amp;quot;typing (esc)-M&amp;quot; or even both: &amp;quot;setting the mark by typing (esc)-M&amp;quot;. By viewing the referent in terms of the generation relationship, we can claim that &amp;quot;U typing (esc)-M at a given time generates U setting the mark at that given time&amp;quot;. This relationship allows us to refer to both or either of the referents without having to make a distinction between the two of them. Note that generation is a relationship between the specific events, the event-tokens and not the types.</Paragraph>
    <Paragraph position="14"> Enablement, as opposed tO generation, has been described as follows: if an agent performs an action and thereby puts the world in a state in which a second action will generate a third action, then we can say that the agent's performance of the first action enables the generation of the third by the second. For example, (18) U: I want to move the cursor 20 characters to the right. How can I do it? E: Like all other commands in EMACS, these commands can be given arguments which cause them to be executed repeatedly. The way you give a command a repeat count is by typing control-U and then the digits before you type the command. For instance, control-U 20 (RIGHT-ARROW} moves forward 20 characters.</Paragraph>
    <Paragraph position="15"> Go ahead, try that.</Paragraph>
    <Paragraph position="16"> In .this example, the pronoun that refers to the entire enablement relationship: &amp;quot;U typing control-U enables a given number (20) and the RIGHT-ARROW key to be typed which in term generates the cursor to move 20 characters forward&amp;quot;. U typing control-U is viewed as an enablement relationship: if U does not type control-U and only types 20 followed by the RIGHT-ARROW, U's desired goal of making the cursor move 20 characters forward will not be accomplished.</Paragraph>
    <Paragraph position="17"> To summarize, for each representation of the individual events characterized by each sentence, we can relate the pronoun to the event, to a sequence of events, or to events related by either the generation or enablement relationships.</Paragraph>
  </Section>
class="xml-element"></Paper>